Our Corrections Had Nowhere to Go

This site keeps a public lab. Products that did not make it, full research breakdowns, and a running record of how the system that judges them changes. That record is split into three shelves: corrections to things we published, rules of ours that changed, and capabilities the system gained.

Two of those three shelves held one post each. One of them had held that single post since January.

They were not short of events. They were short of a question. The law that files one of these entries asked whether a change was worth publishing, and it never asked which shelf the entry belonged on, so every entry landed on the same one. The proof was already sitting there in plain sight: a post about a claim of ours that turned out to be wrong, and a post about a judgment call on a product we had rejected and then approved, both filed as general system news.

So this was never a writing problem. It was a filing problem, and filing problems are fixed with a question, not with output.

The filing law now asks four things, in order, at the moment a change lands. Did something we published turn out to be wrong. Did a rule or threshold of ours change. Did the system gain something it could not do before. If none of those, the change is internal and stays internal. The first one that fits decides, and a correction outranks the other two, because an error we published is the most useful thing we can tell you about ourselves.

The description of the corrections shelf changed with it. It used to say corrections arrive when a maker verifies something for us. That was true, and it was also the smallest part of the truth: most of the corrections on this site were found by our own checks running against our own published claims, and the wording we published had no room for them. It does now.

Two more things were found while measuring, and both belong here rather than in a private note.

The file this site publishes for AI systems listed our lab shelves twice. One list named all of them. The other, the one that carries the actual web address of each shelf, was missing the corrections shelf entirely. A machine reading for our corrections record found the promise and never found the door. The structured data we publish for search engines had the same hole.

And the code that sets those shelf descriptions only ever ran when a shelf was created. Editing a description in the code did nothing on a site where the shelf had existed for months. The code and the live site could disagree indefinitely, and nothing was watching for it. The descriptions are now synced to the code whenever the code says they changed.

Finally, the two pages that show you how this system evolves were reading one shelf out of three. A record of our corrections that no page points at is a record that may as well not exist. Both pages read all three now, and each shelf has its own link, so you can go straight to the one you came for.

Later note, 21 August 2026: the corrections shelf this entry created carries one more obligation at that same moment, and for a while it was not being met. Every correction is supposed to record which of two things it touched, and nine of ours recorded nothing. We wrote down why, and what now runs at the moment a correction is written.

Every change here was made by one AI working the passes and the laws, and a human approving the gate before it shipped.